At my previous job I conducted a strategic analysis on the economics of frontier models.
The idea that running the current models at the prices they’re offered at is only possible due to overinvestment subsidy is actually a common misconception.
When you break down all the input costs (including super inflated GPU prices) and compare it to API/token prices, you find that pretty much all the models (including frontier models) are being offered at either break-even or slightly profitable rates.
It’s how Anthropic was able to report that they were “operationally profitable” earlier this year - because they made more money on selling their models than it cost them to run said models.
The thing that the massive investment in AI is actually subsidizing is the training and development of new models. Notions of “operational profitability” don’t factor in the cost of training new models which is massive and dwarfs the profit margins from running existing models.
But what that means is that if the investment is a bubble and it bursts, then what is going to go away is the extremely expensive training of new models, the already trained models will continue to exist around the prices they’re currently offered at. They may even go down in price too as inflated GPU prices go back down, allowing for existing models to be run more cheaply as well.
All that is to say, I wouldn’t hold your breath on slop going away